Case study: The arithmetic-geometric means inequality

Mathematical results are not just inert facts, but can live in a variety of different
neighbourhoods. This is not generalluy evident from the mere statement of the result,
but is likely to be seen from the proof. To prove a result is to first adopt a point of view
towards the result and to put it into some context, from which the strategy of the proof
will be constructed.
A nice example of this is the arithmetic-geometry means inequality. We will discuss
various arguments for this and what they convey to us about the result. Recall that the
inequality is

ab

1
(a + b)
2

(1)

whenever a and b are nonnegative real numbers.

1. Meaning and interpretation. The inequality (1) is given in algebraic terms. However, putting side the case where one of a and b might vanish (which is obvious), we can
express in a more informal way: the geometric mean of two positive numbers does not
exceed their arithemtic mean. The geometric mean g of a and b is that positive number
g for which a/g = g/b; the arithmetic mean is that number m for which a m = m b.
Then we have to argue that g m.
We can view the situation in this way. Without losing generality, let us suppose that
a b. The condition for the geometric mean means that the factor that we multiply a by
to get g is the same as the factor we multiply g by to get b. Since a is larger than g, this
means that we have to come down further from a to get to g than we have to go from
g to b. Thus, going down to g takes us more than halfway between a and b, and so below
m, the exact halfway point.
The foregoing intuitive and informal explanation may leave much to be desired from
the point of view of a properly composed proof, but it exploits the semantic content of the
terms involved in a way that the more formal arguments given later fail to do. We tackle
the inequality at the level of its meaning, and the role of this argument is to bring to the
front of our minds the essence of the result.
2. The straightforward technical argument. A general strategy for establishing inequalities is to look at the difference of the members on both sides of the inequality and
thus convert it to a relationship between some quantity and 0. In order to avoid the
1

awkwardness of the surd, we establish the equivalent inequaliy

4ab (a + b)2 .
Subtracting the left member from the right, we obtain the quantity
(a + b)2 4ab = a2 2ab + b2 = (a b)2

(2)

which is always greater than or equal to 0, with equality occurring only if a = b.

While the details are specific to the inequality, the strategy is a general one and places
the inequality is a class of similar polynomial inequalities. For example, the generalization
of this to three numbers is the following:

abc

1
(a + b + c) ,
3

for nonnegative values of a, b, c. The substitutions x = u3 , y = v 3 , z = w3 yields the

equivalent
3uvw u3 + v 3 + w3 .
The strategy works on this; just note that
u3 + v 3 + w3 3uvw = (u + v + w)(u2 + v 2 + w2 uv vw wu)
1
= (u + v + w)[(u v)2 + (v w)2 + (w u)2 ] 0 .
2
At this stage, the proof generates some questions for investigation. The analogue of
these inequalities for n variables amounts to
nu1 u2 un un1 + un2 + + unn
for nonnegative values of the ui , and we might ask whether a similar strategy works to
establish this, viz, to write the difference
un1 + + unn nu1 u2 un
as a product of polynomials each of which is clearly seem to be nonnegative for some
elementary reason.
Observe that this argument is a contrast to the first. While the first brings out the
interpretation of the means and their relationship, this one is completely technical and so
2

becomes an exemplar for what might be done in similar situations. It does not require the
understanding of the situation that the first argument evokes.
3. The theory of the quadratic. The proof that one might adduce for a result depends
very much on what one sees as the significant aspects of the situation. Argument 1 saw
that significance in the origins of the means an Argument 2 saw the possibility of taking
a difference that could be shown as positive. Here we observe that there is an interplay
between the sum and the product of two quantitites, which reminds us of a quadratic
polynomial whose coefficients can be described in terms of the sum and product of their
roots. Indeed, consider the quadratic equation:
0 = (x a)(x b) = x2 (a + b)x + ab .
It (clearly) has real roots, and so its discriminant must be nonnegative:
(a + b)2 4ab 0 .
But, of course, this is precisely what we are looking for!
4. A geometric model. How we see a mathematical result is reflected in how we model
it in our minds, and this can be evoked in the proof. For example, consider a circle with
diameter AB of length a + b. Determine the point P on the diameter for which AP = a
and P B = b, and draw a perpendicular to AB that meets the circumference of the circle
at Q. Appealing to the similar triangles AP Q and QP B, we see that A : P Q = P Q : QB

whereupon P Q ab. But P Q is no greater than the radius of the circle (in particular,
that radius to which it is parallel), and the result follows.
5. Concavity. Anyone with a structural understanding of the arithmetic operations
will observe that the geometric and arithmetic means are analogues with multiplication
and exponentiation in the former playing the roles of addition and multiplication in the
latter The conduit of this relationship is the logarithm function. Accordingly, in the case
that a and b are both positive, we see that the arithmetic-geometric means inequality is
equivalent to
1
1
(log a + log b) log( )(a + b) .
2
2
If we know, for some reason, that the logarithm function is concave (i.e. that if satisfies
the condition
f (ta + (1 t)b) tf (a) + (1 t)f (b)
whenever 0 t 1), then this inequality, being a special case, is immediate.
3

The validity of this proof depends on what we are permitted to assume at this stage.
If we need the arithmetic-geometric means inequality to establish the concavity of the logarithm, then the argument is circular, and all we have done is to state a connection between
two properties that may or may not be true. However, if we can establish independently
the concavity of the logarithm function (say by noting that its second derivative is negative), then we not only have a neat structural argument for the inequality but, returning
by the route by which we came, a bounty of additional results:
at b1t ta + (1 t)b
whenever a, b > 0 and 0 t 1. Because this holds for arbitrary exponents t in the closed
unit interval, we have gotten a result that is beyond the realm of algebra and is the fruit
of analysis.
This argument opens up other possible areas of investigation. The logarithm function
is not the only concave function; they abound. Each such function generates its own
inequality. Some of these may be mundane and devoid of interest or application, but
others may have wide-ranging significance.
6. An abstract approach. While the foregoing argument leads us to a level of generality, a more acute use of the logarithm function and of the perception that the logarithm
of the geometric means is the arithmetic mean of the logarithms leads us to an argument
of astonishing simplicity and generality. The price we pay for this is an abstractness,
which may drain the result of some of its immediate semantic content but crystallizes its
structural aspects.
Let S be an arbitrary set an let R denote the system of real numbers. For each
function f from S to R, we suppose that an average A(f ) is defined that satisfies these
five axioms:
(i) A(f ) is a real number;
(ii) A(cf ) = cA(f ) fpr amy real constant c;
(iii) A(f + g) = A(f ) + A(g) for any two functions f and g;
(iv) A(f ) 0 for all functions f what assume only nonnegative values;
(v) A(1) = 1, where 1 denotes the function on s that everywhere assumes the value 1.
In other words, A is a positive normalized linear functional on the vector space of real
functions from S to R. Let us call A(f ) the arithmetic mean of f . It is a consequence of
(iii) and (iv) that
4

(vi) A(f ) A(g) whenever f g (i.e. f (s) g(s) for each s S. This is the condition
of monotonicity.
If f is a function on S that is everywhere positive, then log f , the natural logarithm
of f is everywhere defined. We define the geometric mean G(f ) by
G(f ) = exp A(log f )
where exp x = ex , e being the base of natural logarithms.
We wish to establish that
G(f ) A(f )

(4)

whenever f is a strictly positive function defined on S. To avoid needless complications,

we will confine our considerations to the situation where A(f ) > 0.
For example, if we let S be a set with exactly two points, then each real function on
S can be represented by a pair (a, b) of real numbers, and we can take A(a, b) = 21 (a + b).
It is a consequence of the concavity of the logarithm function as well as a simple
observation from the graph of the logarithm that log t t 1 for t > 0. Let s be any point
of S and apply this inequality to t = f (s)/A(f ). Then, by monotonicity,

log

f
f

1
A(f )
A(f )

since this holds whenever both sides are evaluated at any particular value of s.
From the characteristic property of the logarithm, this inequality can be converted to
log f log A(f )1

f
1 .
A(f )

(5)

(Recall that A(f ) is a real number. We regard this as an inequality between two functions.
Taking the average of the left side and applying properties (ii), (iii) and (v) of A, we
find that
A(log f log A(f )1)
= A(log f ) A(log A(f )1
= A(log f ) (log A(f ))A(1) = A(log f ) (log A(f )) .
5

Taking the average of the right side and applying properties (ii) and (v), we get that
A(f /A(f ) 1) = A(f /A(f )) A(1) = (1/A(f ))A(f ) 1 = 1 1 = 0 .

Thus, (5) yields A(log f ) (log A(f )) 0, or

A(log f ) log A(f ) .
Exponentiating both sides of this inequality leads to the desired G(f ) A(f ).
This argument transports us to a different world, and may be difficult for someone to
follow who has not already had experience in the more abstract realms of mathematics.
At each stage, one may have to consciously reflect on exactly what kind of quantity each
of the symbols represents and in order to ascertain that the operations carried out are
properly sited. The arithmetic-geometric means inequality that we started out with is a
picayune application of this result. More generally, we can apply it to an operator defined
on a space of real functions defined on a set of n points; such a function can be represented
by a vector f = (a1 , a2 , , an ) displaying its respective values.
Consider a vector (w1 , w2 , , wn ) of weights that satisfy the two conditions that each
wi is nonnegative and that w1 + w2 + + wn = 1. Then we can define the weighted
average
A(f ) = w1 a1 + w2 a2 + + wn an .
The general result that has just been obtained yields
wn
1 w2
aw
1 a2 an w1 a1 + w2 a2 + + wn an .

We have here six different arguments of the arithmetic-geometric means inequality,

and each of them conveys a different flavour of the inequality. Some of them reach out
to related areas of mathematics, enriching these other areas with an increased sense of
relevance, generality and value, while others plough into the very meaning that underlies
the symbols.